Why Crashes Happen Here - A Combination Model of Land Use and Road Network

Key Workds:

Crash risk, Land use, Road network, Machine learning

Research Background and Motivation:

With accelerating urbanization and motorization, traffic safety has become a common challenge for cities worldwide. While traditional studies have focused on driver behavior(1), vehicle characteristics(2), or local roadway geometry(3), a growing body of evidence highlights the foundational role of land use(4) patterns and road network(5) structure in shaping traffic conflicts and crash risk. At the same time, network topology determines how traffic is distributed, which links become structural bottlenecks or unavoidable corridors, and where crash hotspots tend to emerge.

This research topic is to define the affection of land use and road network to crashes.

(1)Ma, M., Yan, X., Huang, H., & Abdel-Aty, M. (2010). Safety of Public Transportation Occupational Drivers: Risk Perception, Attitudes, and Driving Behavior. Transportation Research Record: Journal of the Transportation Research Board, 2145(1), 72-79. https://doi.org/10.3141/2145-09 (Original work published 2010)

(2)Metzger, K. B., Sartin, E., Foss, R. D., Joyce, N., & Curry, A. E. (2020). Vehicle safety characteristics in vulnerable driver populations. Traffic Injury Prevention, 21(sup1), S54–S59. https://doi.org/10.1080/15389588.2020.1805445

(3)Reagan J A. Designing for safety by analyzing road geometric[J]. Public Roads, 1994, 63(2): 21-27.

(4)Pulugurtha S S, Duddu V R, Kotagiri Y. Traffic analysis zone level crash estimation models based on land use characteristics[J]. Accident Analysis & Prevention, 2013, 50: 678-687.

(5)Wang X, Wu X, Abdel-Aty M, et al. Investigation of road network features and safety performance[J]. Accident Analysis & Prevention, 2013, 56: 22-31.

Direction of Research - land use

The part of land use is below, it is combine the python and story map.

https://storymaps.arcgis.com/stories/6dfb2e4b4d314008acafb68f1cffd10c

This section mainly demonstrates how land use influences traffic crashes. It identifies the spatiotemporal patterns of crash occurrence and further examines how land-use functions and land-use transitions may affect the likelihood of crashes.

Direction of Research - road network

	CRN	ARRIVAL_TM	AUTOMOBILE_COUNT	BELTED_DEATH_COUNT	BELTED_SUSP_SERIOUS_INJ_COUNT	BICYCLE_COUNT	BICYCLE_DEATH_COUNT	BICYCLE_SUSP_SERIOUS_INJ_COUNT	BUS_COUNT	CHLDPAS_DEATH_COUNT	...	WORK_ZONE_LOC	WORK_ZONE_TYPE	WZ_CLOSE_DETOUR	WZ_FLAGGER	WZ_LAW_OFFCR_IND	WZ_LN_CLOSURE	WZ_MOVING	WZ_OTHER	WZ_SHLDER_MDN	WZ_WORKERS_INJ_KILLED
0	2020029257	1957.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
1	2020008631	942.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
2	2020006834	700.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
3	2020006451	825.0	2	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
4	2020008695	1115.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
45585	2025027435	510.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
45586	2025012535	NaN	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
45587	2025043500	2031.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
45588	2025007815	NaN	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
45589	2025026159	238.0	0	0	0	0	0	0	1	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN

45590 rows × 99 columns

This figure illustrates the integration of traffic crash data for Philadelphia from 2020 to 2024. Crash records from multiple CSV files were read and concatenated into a single dataset, resulting in a comprehensive database with approximately 46,000 observations and 99 variables. The dataset includes detailed information on crash timing, involved modes, and injury outcomes, providing a solid foundation for subsequent spatiotemporal and network-based analyses.

This step extracts the unified boundary of Philadelphia using planning district polygons. By dissolving individual planning districts into a single geometry, a consistent citywide boundary is obtained. This boundary is used to constrain both road network extraction and crash analysis, ensuring spatial consistency across all analytical steps.

(<Figure size 800x800 with 1 Axes>, <Axes: >)

Using the extracted city boundary, a drivable road network for Philadelphia was downloaded from OpenStreetMap via OSMnx. The resulting network, shown after projection, captures the spatial structure and connectivity of the city’s roadway system. This network serves as the base for linking crash occurrences to specific road segments.

			osmid	oneway	name	highway	reversed	length	lanes	maxspeed	service	geometry	ref	bridge	access	tunnel	width	junction
u	v	key
109726936	109726940	0	12108955	True	Brunner Street	residential	False	174.382	NaN	NaN	NaN	LINESTRING (-75.15542 40.01863, -75.15717 40.0...	NaN	NaN	NaN	NaN	NaN	NaN
	109992543	0	43226669	False	Germantown Avenue	primary	False	55.074	2	25 mph	disused_tram	LINESTRING (-75.15542 40.01863, -75.15569 40.0...	NaN	NaN	NaN	NaN	NaN	NaN
	109992535	0	43226669	False	Germantown Avenue	primary	True	7.178	2	25 mph	disused_tram	LINESTRING (-75.15542 40.01863, -75.15538 40.0...	NaN	NaN	NaN	NaN	NaN	NaN
109726940	109726950	0	302956448	False	Wayne Avenue	residential	False	11.448	NaN	NaN	NaN	LINESTRING (-75.15717 40.01782, -75.15710 40.0...	NaN	NaN	NaN	NaN	NaN	NaN
	110047495	0	302956448	False	Wayne Avenue	residential	True	34.627	NaN	NaN	NaN	LINESTRING (-75.15717 40.01782, -75.15738 40.0...	NaN	NaN	NaN	NaN	NaN	NaN

In this step, the road network is converted into a GeoDataFrame containing only edge-level information. Each road segment includes attributes such as OpenStreetMap ID, road name, functional class (e.g., residential, primary), directionality, segment length, and number of lanes. These attributes enable analysis of how roadway characteristics relate to crash occurrence.

	CRN	ARRIVAL_TM	AUTOMOBILE_COUNT	BELTED_DEATH_COUNT	BELTED_SUSP_SERIOUS_INJ_COUNT	BICYCLE_COUNT	BICYCLE_DEATH_COUNT	BICYCLE_SUSP_SERIOUS_INJ_COUNT	BUS_COUNT	CHLDPAS_DEATH_COUNT	...	WORK_ZONE_TYPE	WZ_CLOSE_DETOUR	WZ_FLAGGER	WZ_LAW_OFFCR_IND	WZ_LN_CLOSURE	WZ_MOVING	WZ_OTHER	WZ_SHLDER_MDN	WZ_WORKERS_INJ_KILLED	geometry
0	2020029257	1957.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.09453 39.99463)
1	2020008631	942.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.00614 40.04026)
2	2020006834	700.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.16127 40.02431)
3	2020006451	825.0	2	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-74.99057 40.11199)
4	2020008695	1115.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.20008 40.00740)
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
45585	2025027435	510.0	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.17294 39.92485)
45586	2025012535	NaN	1	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.11726 40.03163)
45587	2025043500	2031.0	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.15036 39.99339)
45588	2025007815	NaN	0	0	0	0	0	0	0	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.15567 40.02052)
45589	2025026159	238.0	0	0	0	0	0	0	1	0	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	POINT (-75.11345 39.99644)

45405 rows × 100 columns

This figure shows the conversion of crash records from tabular data into spatial point features using reported latitude and longitude coordinates. After removing records with missing location information, crash points are stored as a GeoDataFrame in the WGS84 coordinate system (EPSG:4326). This step enables spatial linkage between crash events and the road network.

(45331, 100)

(15500, 3)

This step assigns each crash event to its nearest road segment in the network. First, the unioned road geometry is used to define the effective city boundary, and crash points are filtered to retain only those located within this boundary. Both the road network and crash points are then projected to a planar coordinate system (EPSG:2272) to ensure accurate distance calculations. Using OSMnx’s nearest edge function, each crash point is matched to the closest road segment based on spatial proximity. The corresponding node pairs (u, v) are recorded for each crash, allowing crashes to be aggregated at the road-segment level. This results in an edge-based crash count dataset, which forms the basis for subsequent calculation of crash intensity and risk metrics.

	u	v	key	osmid	oneway	name	highway	reversed	length	lanes	maxspeed	service	geometry	ref	bridge	access	tunnel	width	junction	crash_count
0	109726936	109726940	0	12108955	True	Brunner Street	residential	False	174.382	NaN	NaN	NaN	LINESTRING (-75.15542 40.01863, -75.15717 40.0...	NaN	NaN	NaN	NaN	NaN	NaN	1.0
1	109726936	109992543	0	43226669	False	Germantown Avenue	primary	False	55.074	2	25 mph	disused_tram	LINESTRING (-75.15542 40.01863, -75.15569 40.0...	NaN	NaN	NaN	NaN	NaN	NaN	4.0
2	109726936	109992535	0	43226669	False	Germantown Avenue	primary	True	7.178	2	25 mph	disused_tram	LINESTRING (-75.15542 40.01863, -75.15538 40.0...	NaN	NaN	NaN	NaN	NaN	NaN	0.0
3	109726940	109726950	0	302956448	False	Wayne Avenue	residential	False	11.448	NaN	NaN	NaN	LINESTRING (-75.15717 40.01782, -75.15710 40.0...	NaN	NaN	NaN	NaN	NaN	NaN	0.0
4	109726940	110047495	0	302956448	False	Wayne Avenue	residential	True	34.627	NaN	NaN	NaN	LINESTRING (-75.15717 40.01782, -75.15738 40.0...	NaN	NaN	NaN	NaN	NaN	NaN	1.0

Road segment data are merged with crash counts using node pairs (u, v) as unique identifiers. Crash counts are assigned to each road segment, with segments experiencing no crashes set to zero. This produces a complete edge-level dataset that directly associates roadway characteristics with crash frequency.

This histogram presents the distribution of crash index values across road segments with nonzero crash occurrences. The distribution is approximately unimodal, with most values concentrated in the mid-range and a noticeable right-skewed tail. This pattern indicates that while most road segments experience moderate crash risk, a small subset exhibits disproportionately high risk, highlighting spatial inequality in crash exposure.

Make this Notebook Trusted to load map: File -> Trust Notebook

This interactive map visualizes the spatial distribution of the crash index across road segments with nonzero crash occurrences. Road segments are symbolized using a continuous color scale, where darker colors indicate higher crash index values. A dark basemap is applied to enhance visual contrast, and road segments are displayed with increased line weight to improve readability. Interactive tooltips provide detailed information for each segment, including node identifiers (u, v) and the corresponding crash index value. The map reveals clear spatial heterogeneity in crash risk, with higher crash index values clustering along major corridors and dense urban areas, suggesting that crash risk is unevenly distributed across the road network.

u          v          key
109726936  109726940  0      174.382
           109992543  0       55.074
           109992535  0        7.178
109726940  109726950  0       11.448
           110047495  0       34.627
Name: length, dtype: float64

This figure presents the spatial distribution of road intersection complexity, measured by the mean node degree (degree_mean) associated with each road segment. Higher values indicate segments connected to intersections with greater numbers of converging roads, reflecting more complex intersection structures. The map reveals pronounced spatial variation in intersection complexity across Philadelphia, with higher values concentrated in the dense urban core and along major corridors, while lower values are more common in peripheral and residential areas. This pattern suggests that network structure and connectivity vary substantially across the city, potentially influencing traffic flow, driver behavior, and crash risk.

This figure maps the spatial distribution of local closeness centrality, measured as the mean closeness value associated with each road segment. Higher values indicate segments that are, on average, more centrally located within the local road network and can be reached more efficiently from surrounding nodes. The map shows elevated local closeness in the urban core and along major connecting corridors, while peripheral areas exhibit lower values. This pattern reflects underlying variations in network accessibility and connectivity, which may influence traffic volume, route choice, and, ultimately, crash occurrence.

This figure illustrates the spatial distribution of intersection density, calculated as the number of intersections within a 100-meter buffer around each road segment. Higher values indicate areas with denser intersection patterns and more frequent roadway crossings. The map shows that intersection density is highest in the urban core and in older, fine-grained street networks, while lower densities are observed in peripheral areas with more hierarchical or suburban street layouts. This spatial variation reflects differences in network granularity and may be associated with increased conflict points, potentially contributing to higher crash risk in densely connected areas.

residential      42343
tertiary          8538
primary           4590
secondary         3888
trunk              810
unclassified       536
motorway_link      328
primary_link       167
motorway           158
trunk_link         119
Name: highway_clean, dtype: int64

This figure presents the spatial distribution of road functional types across the study area, based on the OpenStreetMap highway classification. Different colors represent distinct road categories, including residential streets, primary and secondary roads, motorways, and their corresponding link roads. Residential streets dominate the network in terms of total length, forming a dense grid within the urban core, while higher-order roads such as primary roads, trunks, and motorways structure the citywide connectivity and facilitate long-distance travel. This classification captures the hierarchical organization of the road network and provides essential context for examining how crash risk varies across different roadway functions and traffic environments.

This figure compares the spatial distribution of one-way and two-way roads across the study area. One-way streets are primarily concentrated in the dense urban core, where grid-like street patterns and traffic management strategies are more prevalent, while two-way roads dominate in peripheral and lower-density areas. This spatial pattern reflects differences in urban form and traffic control practices and may influence vehicle flow, turning movements, and conflict points. As a result, road directionality represents an important structural characteristic that may contribute to variation in crash occurrence across the road network.

This figure illustrates the spatial variation in road segment length across the study area, with segment lengths capped at 200 meters to reduce the influence of extreme values. Shorter segments are more prevalent in the dense urban core, reflecting fine-grained street grids and frequent intersections, while longer segments are more common in peripheral areas with larger block sizes and more continuous roadway layouts. Road segment length captures an important aspect of network geometry that may influence vehicle speeds, stopping frequency, and driver behavior, and thus represents a relevant factor in explaining variation in crash occurrence across the road network.

count    61661.000000
mean       109.253598
std        107.908745
min          1.901000
50%         83.300000
90%        190.169000
95%        250.012000
99%        488.773000
max       3222.776000
Name: length, dtype: float64

This figure depicts the spatial distribution of posted speed limits across the road network, measured in kilometers per hour (kph). Lower speed limits are predominant within the dense urban core, reflecting traffic-calming measures and high levels of pedestrian and vehicular interaction, while higher speed limits are mainly observed along major arterial roads, highways, and peripheral corridors. This spatial pattern highlights the hierarchical nature of the road network and differences in traffic operating environments. Posted speed limit is a key roadway attribute that influences vehicle speed, crash frequency, and crash severity, and therefore represents an important explanatory variable in the analysis of crash risk.

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
Cell In[3], line 1
----> 1 merged = edges_nonzero.merge(
      2     edges_fix[['u', 'v', 'key', 'degree_mean', 'closeness_mean',
      3                'intersection_density', 'speed_kph', 'length']],
      4     on=['u', 'v', 'key'],
      5     how='left'
      6 )

NameError: name 'edges_nonzero' is not defined

This figure presents the Pearson correlation matrix between the crash index and key roadway and network characteristics. The crash index shows a moderate positive correlation with mean closeness centrality (r = 0.29), suggesting that road segments with higher local accessibility tend to experience higher crash risk. In contrast, crash index is negatively correlated with road segment length (r = −0.49), indicating that shorter segments—often associated with denser intersection patterns—are linked to higher crash occurrence. Correlations between crash index and other variables, including degree mean, intersection density, and speed limit, are relatively weak, highlighting the complex and potentially nonlinear relationships between network structure and crash risk. Additionally, moderate correlations among explanatory variables (e.g., between closeness centrality and intersection density) suggest some shared structural information but do not indicate severe multicollinearity, supporting their joint inclusion in subsequent modeling.

	variable	VIF
0	degree_mean	10.870226
1	closeness_mean	7.775618
2	intersection_density	10.121696
3	speed_kph	24.624842
4	length_x	3.461989

This table reports the Variance Inflation Factor (VIF) values for key roadway and network variables included in the analysis. Several variables exhibit elevated VIF values, most notably posted speed limit (VIF = 24.62), degree mean (VIF = 10.87), and intersection density (VIF = 10.12), indicating substantial multicollinearity among these predictors. Closeness centrality also shows moderate multicollinearity (VIF = 7.78), while road segment length displays a relatively low VIF (3.46). These results suggest that certain network structure and roadway design variables capture overlapping information, likely reflecting shared underlying spatial and functional characteristics of the road network. As a result, caution is required when jointly including these variables in regression models, and additional modeling strategies—such as variable selection, regularization, or nonlinear methods—may be necessary to ensure stable and interpretable estimates.

                            OLS Regression Results                            
==============================================================================
Dep. Variable:            crash_index   R-squared:                       0.262
Model:                            OLS   Adj. R-squared:                  0.262
Method:                 Least Squares   F-statistic:                     1819.
Date:                Fri, 12 Dec 2025   Prob (F-statistic):               0.00
Time:                        00:16:18   Log-Likelihood:                 15123.
No. Observations:               15383   AIC:                        -3.024e+04
Df Residuals:                   15379   BIC:                        -3.021e+04
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
========================================================================================
                           coef    std err          t      P>|t|      [0.025      0.975]
----------------------------------------------------------------------------------------
const                    0.4362      0.002    180.756      0.000       0.431       0.441
closeness_mean           0.0063      0.000     12.820      0.000       0.005       0.007
intersection_density     0.0086      0.001     10.481      0.000       0.007       0.010
length_x                -0.0005   8.04e-06    -59.314      0.000      -0.000      -0.000
==============================================================================
Omnibus:                     3137.456   Durbin-Watson:                   1.855
Prob(Omnibus):                  0.000   Jarque-Bera (JB):             6147.405
Skew:                           1.238   Prob(JB):                         0.00
Kurtosis:                       4.861   Cond. No.                         550.
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

RandomForestRegressor(max_depth=12, min_samples_leaf=20, n_estimators=300,
                      random_state=42)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

This figure presents partial dependence plots illustrating the marginal effects of selected roadway and network variables on predicted crash risk, based on the random forest model. The plot for mean closeness centrality shows a relatively flat relationship, with crash risk remaining fairly stable across most values, suggesting a limited marginal effect once other variables are controlled. Intersection density exhibits a modest positive association with crash risk, indicating that segments located in denser intersection environments tend to experience slightly higher predicted risk. In contrast, road segment length displays a strong negative relationship with crash risk, with shorter segments associated with substantially higher predicted crash risk. This pattern is consistent with the concentration of conflict points and frequent stopping behavior in short, fine-grained street segments. Overall, these plots highlight nonlinear relationships and heterogeneous effects that are not easily captured by linear models, underscoring the value of machine learning methods for understanding crash risk mechanisms.

This figure shows the mean absolute SHAP values for selected predictors, representing their average contribution to the crash risk predictions of the random forest model. Road segment length emerges as the most influential variable by a substantial margin, indicating that variation in segment length plays a dominant role in shaping predicted crash risk. Intersection density and mean closeness centrality exhibit considerably smaller but non-negligible impacts, suggesting that local network structure and accessibility also contribute to crash risk, albeit to a lesser extent. These results reinforce the findings from the partial dependence analysis, highlighting the importance of roadway geometry and network granularity in explaining crash risk, and demonstrating the value of SHAP-based interpretation for identifying key drivers in nonlinear models.

This SHAP summary plot illustrates the distribution of feature impacts on the crash risk predictions across all observations. Road segment length shows the widest spread of SHAP values, indicating substantial heterogeneity in its influence on crash risk. Shorter segments (low values, shown in blue) are associated with positive SHAP values, meaning they increase predicted crash risk, whereas longer segments (high values, shown in red) tend to reduce predicted risk. Intersection density exhibits a smaller but consistent positive effect, with higher densities contributing modestly to increased crash risk. Mean closeness centrality displays a relatively narrow distribution around zero, suggesting a weaker and more context-dependent influence on crash risk. Overall, the SHAP results highlight road geometry and local network structure as primary drivers of crash risk, while also revealing nonlinear and asymmetric effects that are not captured by linear modeling approaches.

This SHAP dependence plot illustrates the relationship between road segment length and its contribution to predicted crash risk, with points colored by mean closeness centrality. Short road segments are associated with strongly positive SHAP values, indicating a substantial increase in predicted crash risk, while SHAP values decline sharply as segment length increases. Beyond a certain length threshold, the marginal effect of additional length becomes relatively stable and remains slightly negative. The color gradient suggests an interaction effect, where segments with higher closeness centrality tend to exhibit higher crash risk at shorter lengths. This pattern highlights a nonlinear relationship between road geometry and crash risk, emphasizing that short, highly connected road segments are particularly vulnerable to elevated crash risk.

Conclusion

This study examined the relationship between traffic crash risk and roadway network characteristics using a network-based, multi-method analytical framework. By integrating multi-year crash data with detailed road network attributes, the analysis moved beyond point-based crash mapping to an edge-level understanding of how roadway geometry and network structure shape crash risk across space.

The results reveal substantial spatial heterogeneity in crash risk across the road network. Crash index values are unevenly distributed, with higher-risk segments clustering in dense urban areas and along major corridors. Exploratory correlation analysis suggests that crash risk is moderately associated with network accessibility and strongly related to road segment length, while linear associations with other variables remain relatively weak. Variance inflation factor diagnostics further indicate notable multicollinearity among several network measures, highlighting the limitations of relying solely on linear regression approaches to interpret crash mechanisms.

Machine learning–based interpretation provides clearer insights into the underlying relationships. Partial dependence plots and SHAP analyses consistently identify road segment length as the dominant predictor of crash risk. Short road segments—often corresponding to dense intersection environments and fine-grained street grids—are associated with substantially higher predicted crash risk, while longer segments tend to reduce risk. Intersection density and closeness centrality play secondary but meaningful roles, particularly in combination with short segment lengths, indicating that highly connected and accessible network locations experience elevated exposure to conflicts. These findings underscore the nonlinear and interactive nature of crash risk determinants, which are not fully captured by traditional linear models.

Overall, the findings suggest that crash risk is closely tied to the geometric and structural characteristics of the road network rather than any single factor in isolation. Dense, highly connected urban street environments—while critical for accessibility—also concentrate conflict points and increase crash vulnerability. From a planning and design perspective, this highlights the importance of targeted safety interventions on short, highly connected road segments, such as intersection redesign, traffic calming, and access management. Methodologically, the study demonstrates the value of combining network analysis with interpretable machine learning techniques to better understand complex crash risk mechanisms and to support data-driven transportation safety planning.